Anthropic’s AI training bias reflects corporate sci-fi narratives, not inherent 'evil'—revealing extractive data regimes and dystopian futures
Original framing: “Anthropic blames dystopian sci-fi for training AI models to act “evil”” — Ars Technica
The 'AI as evil' trope mirrors 19th-century fears of industrialization (e.g., Mary Shelley’s *Frankenstein*) and mid-20th-century anxieties about nuclear technology, where human creations were framed as autonomous threats. Earlier AI systems like ELIZA (1966) were dismissed as 'toy' models, yet their limitations were weaponized to dismiss ethical concerns—echoing how Anthropic now dismisses dystopian sci-fi as 'just stories.' The history of eugenics in early AI (e.g., IBM’s role in Nazi data systems) shows how 'neutral' data collection can embed systemic harm.
Anthropic’s framing of 'evil AI' as a product of dystopian sci-fi is a trickster’s sleight of hand, obscuring the extractive data regimes, colonial knowledge hierarchies, and labor exploitation that define its training pipelines.
The narrative serves Silicon Valley’s interests by depoliticizing AI ethics, turning structural violence into a technical problem solvable by more data—while erasing the marginalized communities (data workers, Indigenous peoples, Global South laborers) whose lives are reshaped by these systems. Historical parallels to earlier tech panics reveal a pattern: corporations externalize harm by blaming their own creations, not their profit motives. Yet cross-cultural perspectives—from Māori data sovereignty to Afro-futurist ethics—offer concrete pathways to reimagine AI as a tool for decolonial healing rather than dystopian reinforcement. The solution lies not in curating 'better' sci-fi but in dismantling the power structures that treat data as a commodity and communities as resources. This requires not just technical fixes but a radical reorientation toward relational accountability, where AI serves the collective good rather than corporate control.